[PATCH] D47194: [AMDGPU] Fixed non-uniform addr64 MUBUF in shader

[Tim Renouf via Phabricator via llvm-commits]

tpr added inline comments.


================
Comment at: lib/Target/AMDGPU/SIInstrInfo.cpp:3667
   if (isMIMG(MI) ||
       (AMDGPU::isShader(MF.getFunction().getCallingConv()) &&
+       (isMUBUF(MI) || isMTBUF(MI))
----------------
arsenm wrote:
> I don't like that this was checking the calling convention at all. Can we just eliminate that check altogether?
So (from a compute point of view) you think it is ok for all non-addr64 cases of MUBUF/MTBUF to go in here? I think that is probably ok, but then I'm not sure why Nicolai added the calling convention check in the first place.


================
Comment at: lib/Target/AMDGPU/SIInstrInfo.cpp:3669
+       (isMUBUF(MI) || isMTBUF(MI))
+       && !isAddr64(MI))) {
     MachineOperand *SRsrc = getNamedOperand(MI, AMDGPU::OpName::srsrc);
----------------
arsenm wrote:
> You should be able to avoid adding isAddr64 by just moving the getNamedOperand call for vaddr below up here
I tried that, and I got false positives on offen/idxen/bothen variants.


Repository:
  rL LLVM

https://reviews.llvm.org/D47194